You have a prompt saved somewhere. A block of text you paste into a chatbot every Sunday before you write the newsletter. It names your tone. It lists your sections. It explains who your readers are and what this week is about. You wrote it once, refined it forty times, and you are quietly a little proud of it.
Open it next to last week’s prompt. Read both side by side. Most of the words are identical. The voice instructions match. The section structure matches. The audience description matches. The only part that changed is this week’s stories.
So you have been hand typing the same instructions every week to a task that repeats every week. The industry calls that prompt engineering and sells it as a skill.
A skill with an expiry date that nobody printed on the box.
Is Prompt Engineering a Dying Skill?
The people who build these models keep telling you it is.
Prompt engineering was crowned the job of 2024. Six-figure salaries. Course bundles. LinkedIn badges. Then the models got better at guessing what you meant, and the premium on phrasing started to fall. Salesforce Ben documented the role sliding toward obsolescence through 2025, and quoted Microsoft’s Jared Spataro saying the quiet part out loud: “You don’t have to have the perfect prompt anymore.” The Wall Street Journal reported the same title being phased out across the companies that were hiring for it a year earlier.
The debate hasn’t fully settled. Learn Prompting catalogs the arguments on both sides, and the honest summary is that prompts still matter. Clear input still beats vague input.
And the part that should change how you think about it. The reason prompt engineering is fading is that every model upgrade narrows the gap between a carefully built prompt and a casual request. The skill gets easier to do and less valuable to have with every release.
A skill that gets easier every time the tool improves comes with a countdown that most operators never check.
Compare that to the skills that hold their value. Writing gets harder as you raise your standard. Editorial judgment compounds with every issue you ship. Those are durable. The ability to phrase a request to a machine is a workaround for a machine that can’t yet understand you. The workaround disappears as the understanding improves.
No judgment. The tools trained you to do this. They shipped a blank text box and an implicit instruction to get good at filling it.
Three Interfaces, Sixty Years, One Direction
Step back from newsletters for a second, because the pattern is older than AI.
Jakob Nielsen, who has spent forty years studying how people use computers, counts only three interface paradigms in the entire history of computing. The first was batch processing, around 1945. You specified a complete workflow, handed over a stack of punched cards, and picked up the printout the next morning. The second arrived in the 1960s and ran for sixty years: command based interaction. You issue a command, the computer executes it, you issue the next one. The command line was this. So was every menu and button that followed.
The third paradigm is the one you are living through. Nielsen calls it intent based outcome specification. You state the result you want. The machine works out the steps. You no longer tell the computer how to do the thing. You tell it what you want, and it figures out the how.
Read that arc in one direction. Each paradigm moves more of the work from the human to the machine. Batch made you plan everything in advance. Commands let you correct course one step at a time. Intent lets you skip the steps entirely and name the destination.
Every interface generation since 1945 has moved the same way, taking work off the human and handing it to the machine. Prompting that asks you to specify the steps is walking backward against sixty years of that current.
Now look at your Sunday prompt again. You’re listing steps. Tone, then structure, then audience, then constraints, then this week’s angle. That’s command based interaction wearing the costume of a conversation.
Your saved prompt block is a deck of punched cards with a nicer font. Congrats!
Nielsen goes one step further, and this is the part worth holding onto. He notes that intent based systems still sit below what he called non-command systems back in 1993. A true noncommand system does not even ask you to state intent. The computer acts as a side effect of what you were already doing. Pulling the door handle opens the car. You never issued an open command. You reached for the door, and the lock released on its own.
For a task you repeat every week, that’s the standard worth aiming at. The most evolved interface for your newsletter is the one that needs almost nothing from you, because it already holds the recurring parts. A better prompt box only polishes the paradigm we’re trying to leave.
Why Did the Blank Box Win the Last Decade?
The blank box became the default for a practical reason, and that reason is already expiring.
In 2022, the models could not infer much. A vague request produced a vague result. Phrasing genuinely changed the output, so the people who learned to phrase well got better results, and the skill was real. The blank box was also the cheapest interface to ship. One text field serves code, recipes, birthday toasts, and newsletters with equal indifference. That generality is what carried the chatbot to hundreds of millions of people. The same generality is what leaves it shapeless for any single recurring job.
So newsletter operators inherited a tool built for everyone, plus an unspoken assignment to make it work for one specific publication. They wrote the system prompt. They learned the tricks. They built the saved block. The work was rational at the time.
Two forces have shifted since. The models got better at reading intent, which shrank the payoff on clever phrasing. And tools built for a specific job started holding the recurring parameters, so the blank box stopped being the only option on the table.
Surprising? Only if you have never watched a brand new field quietly get absorbed into the tools around it.
Why Is Prompting the Wrong Interface for a Weekly Task?
There is a precise term for the distance between what you want and what a system makes you do to get it.
Donald Norman named it the gulf of execution in The Design of Everyday Things. The Nielsen Norman Group frames it as the gap between a user’s intentions and the actions the system requires to carry them out. A small gulf means the controls line up with what you meant. A wide gulf means you have to translate your intention into the system’s language before anything happens.
A blank prompt box is a wide gulf of execution. Your intention is simple. You want this week’s issue, in your voice, built from your sources. To get it, you translate that intention into a paragraph of instructions the model can act on. Every Sunday, you perform the translation again. The gulf does not close. It reopens on a weekly schedule.
Good interface design has known how to narrow this for decades. Nielsen’s sixth usability heuristic asks systems to favor recognition over recall: show people their options so they do not have to dredge instructions out of memory. A blank prompt is pure recall. You are remembering and retyping what your newsletter is, from scratch, into an empty field.
There’s one more reason this never resolves on its own. Researchers John Carroll and Mary Beth Rosson described the paradox of the active user: people skip the manual and jump straight to the task. They will not invest upfront time to optimize a system, even when it would save them hours later. So most operators never actually master prompting. They paste a serviceable block, accept a serviceable draft, and move on.
The gulf of execution between you and a blank prompt box reopens 52 times a year, because the interface forgets your newsletter the moment you close the tab.
You’re not a prompt engineer. You’re a writer doing unpaid quality control on a machine’s vocabulary, every single week.
What’s Prompt Fatigue?
Prompt fatigue is the specific drain of describing the same unchanged task to a machine, week after week.
It is different from writer’s block and from the general exhaustion of production. It’s the felt cost of explaining your own newsletter to a machine that should already know it, over and over, on a loop. The parameters that define your publication are stable. Your voice does not reset on Sunday. Your sections do not reshuffle. Your audience does not become strangers. Yet the blank box treats every session like the first one.
For a question you ask once, a long prompt is research. For issue 47 of the same newsletter, it’s data entry you handed to yourself.
This points to a number worth tracking, which is the input-to-output ratio. How much do you have to say to get what you need? For a one-time creative exploration, a rich, detailed prompt is the right trade. You’re paying input to discover an output you can’t yet picture. For a recurring publication, the math inverts. You already know the output you want. The input you keep typing is overhead.
And the perfect prompt was never where the leverage lived anyway. Andrew Ng’s team measured this on the HumanEval coding benchmark. A single zero-shot prompt to GPT 3.5 scored 48.1 percent. The same prompt to the much stronger GPT 4 scored 67.0 percent. Then they wrapped the weaker GPT 3.5 in a structured loop that broke the task into steps, and it reached up to 95.1 percent. The structure around the model added more than two full generations of model upgrades.
Read that gap again.
The takeaway for a newsletter operator is simpler than building an agent. The result comes from how the task is structured, and structure is exactly “the thing” you should never have to rebuild by hand every week. The leverage was never in your phrasing. It lived in the scaffolding around the request, and scaffolding is a one-time job that pays out every issue after.
How Much Should You Have to Tell an AI to Write Your Newsletter?
For a recurring issue, the honest answer is almost nothing.
You should specify the part that changed since last week, and the system should already hold the rest. The voice should be learned from your archive. The sections you always run. The sources you actually read. Stating those again is retyping a setting.
The only input a weekly issue truly needs is the part that changed since last week. Everything else is a setting you keep entering by hand.
This is the design problem HeyNews was built around. The platform learns your voice profile from your published archive once, so the recurring parameters live in the system, off the prompt you used to retype every Sunday. A Select Best button picks candidate stories against your editorial patterns. One click generates the full draft. One click transforms handle shorten, expand, simplify, and similar moves without rewriting instructions each time. Automations can draft a recurring issue on a schedule and hold it for your review, which is about as close to Nielsen’s non-command ideal as a newsletter workflow gets. You named the outcome once. The system produces it on the cadence you set.
That last point matters, so it should be stated plainly. The editorial call stays with you on every draft. The system surfaces a candidate. You decide what ships. I broke down the deeper architecture comparison between a general chatbot and a tool built for this specific job in an earlier post, so this is the short version.
The point holds with or without any particular tool. The interface for a repeating task should ask you for the delta. The full specification is something it can already hold for you.
The Repeated Instruction Audit You Can Run This Week
You don’t need to buy anything to see your own prompt fatigue in numbers. You need your last two AI assisted issues and ten minutes.
Step 1: Capture this week’s input. The next time you generate a draft with any AI tool, copy the full instruction you gave it. Everything. The tone notes, the structure notes, the audience framing, the formatting requests, the angle for the week.
Step 2: Pull last week’s input. Find the equivalent prompt from your previous issue. If you reuse a saved block, pull that and any edits you made on top of it.
Step 3: Mark the repeats. Go line by line. Highlight every instruction that is identical or close to identical across the two issues. The voice direction. The section count. The “keep it punchy” note. The “do not be so formal” correction you make every time.
Step 4: Calculate the ratio. Divide the repeated words by the total. That percentage is the share of your weekly input that carries no new information. It’s the part that a system should already know.
Most operators who run this find that 70 to 90 percent of their prompt is identical week to week. Only the stories and the angle are genuinely new. If your repeat ratio sits above 70 percent, you are hand feeding parameters that belong in a saved profile, and the retyping is the leak.
The audit costs nothing and you can run it without changing tools at all. The number it gives you is the one worth knowing before you decide anything else.
In a Nutshell
- Prompt engineering is a skill that gets easier with every model release, which gives it a built-in shelf life. The skills that compound are writing and editorial judgment.
- Computing has had three interface paradigms in sixty years, and each one moved work from the human to the machine. A prompt that asks you to specify the steps runs against that direction.
- Prompting is a wide gulf in execution for a recurring task. The gap between your intention and the instructions the system requires reopens every week, because the blank box forgets your newsletter each session.
- The leverage was never in the perfect prompt. Ng’s HumanEval data showed a structured loop around GPT 3.5 beating GPT 4 on a single prompt, 95.1 percent to 67.0 percent. Structure is a one time job that keeps paying out every issue.
- You can run the Repeated Instruction Audit this week with no tool: copy two consecutive prompts, mark the lines that repeat, and calculate the share. Above 70 percent means you are retyping settings.
Go back to the saved prompt block from the top of this post. Open it one more time. Count how much of it changes from week to week. The answer is the stories and the angle, and not much else.
So, what to do next? Run the audit on your next issue and get your repeat ratio. Then stop treating the stable 80 percent as something you type and start treating it as something the system should hold. The work that drew you to a newsletter was always the writing and the judgment. The retyping was always just the interface showing its age, caught in an awkward decade between command and intent.
See HeyNews helps you to solve the recurring parts to make them stop being your job: heynews.co
